Researchers from Thinking Machines Lab, in collaboration with Bridgewater AIA Labs, have introduced a new method for training models based on expert judgment to solve complex financial tasks. By using the Qwen3-235B base model with a specialized training recipe, they created a system that significantly outperforms existing frontier models in narrow domains while radically reducing costs.
What Happened
The study demonstrated a specialized pipeline including Interleaved Batching, CISPO loss, and On-Policy Distillation. Applying these methods to the Qwen3-235B base achieved an accuracy of 84.7%, which is 29.8% higher than standard frontier models. Meanwhile, inference costs were reduced by 13.8 times.
Context
The traditional approach to LLM training focuses on massive data collection from the open internet, leading to universal but less accurate models in niche domains. This new approach proposes a shift toward a "differentiated intelligence" paradigm, where the focus moves from parameter volume to the quality of the "signal" through the use of expert judgment and Human-in-the-loop processes.
Why It Matters for the Industry
This case confirms the viability of the vertical AI strategy: creating highly specialized systems is becoming more economically and technically advantageous than using universal giant models. The industry is seeing a shift from the race for parameter scaling (scaling laws) to a race for the quality of expert data and the efficiency of training methods applied to it.
Why It Matters for Users
For businesses, this means the ability to implement high-precision custom solutions for critical processes such as financial analysis, compliance, and risk management. Companies will be able to design efficient internal systems that work more accurately and significantly cheaper than using public APIs from OpenAI or Anthropic.
What Is Not Yet Known / Limitations
There is a critical legal risk regarding intellectual property (IP) issues when using proprietary corporate expertise as a training signal for models.
Sources
Author
Look at AI, Editorial Team